An intersection lane line association method and system based on a pure encoder architecture

By using a Transformer model based on a pure encoder architecture to perform feature encoding and graph convolution processing on lane lines, the accuracy and stability issues of lane line topology association in complex intersection scenarios are solved, and the automated generation of lane line associations in intelligent driving maps is realized.

CN122116681APending Publication Date: 2026-05-29广州祺宸科技有限公司
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610208985.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-13
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Existing technologies struggle to balance the accuracy and stability of lane topology modeling in complex intersection scenarios. Traditional methods fail to explicitly model lane lines as independent nodes, resulting in insufficient consistency and stability of topology relationships.

Method used

A pure encoder-based approach is adopted, which uses a Transformer-based encoder model to preprocess and encode the lane centerlines. Graph convolutional layers are used for forward adjacency feature propagation to generate topological relationships between lane lines, avoiding error accumulation and enhancing explicit modeling capabilities.

Benefits of technology

It outputs more stable and consistent association prediction results in complex intersection scenarios, improves the accuracy of lane line association reasoning and topological information representation capabilities, and supports the automated generation of intelligent driving maps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122116681A_ABST
    Figure CN122116681A_ABST
Patent Text Reader

Abstract

The application belongs to the technical field of intelligent driving map construction, and particularly relates to a kind of intersection lane line association method and system based on pure encoder architecture, the method comprises: obtaining intersection map data containing entering and exiting lane center line;Data preprocessing is carried out to generate initial joint embedding representation;Input into the encoder model based on the architecture of Transformer, the model contains D feature coding networks connected in turn, each network encodes the input feature and outputs the corresponding association probability matrix of the layer;Determine the topological association relationship based on the probability matrix output by the last layer, and generate the intersection topological lane center line.The application directly completes multi-level association relationship prediction in the encoding stage through pure encoder architecture, avoids error accumulation of decoding or autoregressive prediction, and improves the stability and accuracy of lane line topological association prediction under complex intersection.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Road topology prediction method of high-precision map based on Transform

    CN117037095A

  • Intelligent driving online map construction method and system

    CN118071951A